NOTES FROM THE LAB
The journal.
Applied AI, security, research, and the details behind the software.
All posts09
Fingerprints below HTTP: how servers recognise the software that connects
How servers recognise clients by TLS ClientHello, HTTP/2 settings and header order, how defenders use it, where it fails, and what it means for agents.
SecurityLeast privilege for AI agents: put the limits where the model cannot reach
A prompt can ask an agent to behave; roles, tokens, network rules and sandboxes make it. How to bound an agent, and where a human must still confirm.
SecurityPrompt injection is a confused deputy problem
A model reads instructions and data on one channel, so an agent holding authority can be steered by what it reads. The designs that hold by construction.
SecurityWhat an agent harness is made of
An agent is a model plus a harness: loop, tools, context, checks, permissions and state. Why harnesses change scores, and how to compare them.
EngineeringWhen finding bugs gets cheap, fixing becomes the bottleneck
If AI makes vulnerabilities cheap to find, risk is set by triage, fix, release, distribution and deployment. Where the time goes and how to cut it.
SecurityZero-days found by AI systems: what the labs reported and what it costs
A sourced review of zero-days found by AI at Google, OpenAI, Anthropic, Microsoft and others, the patterns behind them, and what defenders should do.
SecurityThe open loop: fast decision models in cloud operations
Small decision models are far cheaper and ten to sixteen times faster than thinking models. Whether that makes monitoring better or worse has not been measured.
Applied AIThe same weights are not the same model
An endpoint is weights plus numerics, decoding, parsers, limits and operations. Each layer can change the answer. How to reason about it, measure it and buy it.
EngineeringBuilding Velofy in the open
Velofy is an applied AI and security lab in Delhi, India. What it builds, why the work is public, and where to start reading.
Applied AI